Performance and resource budget¶
Current authority¶
The original bootstrap limits were superseded by the owner's 2026-10-20 study and resource approval in AUTHORITY. The current local envelope is four CPU threads, 16 GiB RAM, 40 GiB working storage, two hours per job and six hours total, with zero local spend. Licence-reviewed public assets are allowed; HF Jobs are an authorized conditional fallback with a separate bounded job record. Hardware procurement and independent-reviewer contact remain options only. The declaration does not enforce OS limits.
Proposed v1 envelopes¶
These are implementation caps and measurement proposals, not observed throughput or permission to allocate a host:
| Operation | Cap / proposed target | Response on breach |
|---|---|---|
| Prompt and response | Each ≤32,768 UTF-8 bytes; no truncation | LIMIT_EXCEEDED |
| Encoder input | ≤32,768 units; chunks ≤512 units including special tokens | Reject oversize input profile |
| Bundle / each proof | 64 MiB / 16 MiB | Reject before full decode/verification |
| Offline verification | Caller budget ≤60 s total; 2 GiB worker RSS proposed default | Terminate worker, no accepted result |
| Local spool | 32 pending receipts, 1 GiB encrypted bytes, 24 h maximum witness retention proposed default | Stop admitting requests; expire and delete per-receipt key |
| Foreground proof worker | One job at a time; at most one retry within approved total budget | Failed/expired receipt; never increase cap |
| Research cost sample | 5 cold and 30 warm repetitions proposed minimum per admitted length/graph | Incomplete measurement if absent; no extrapolated p95 |
The integer reference targets exactness first. No serving latency or proving time is invented as a demonstrated capability. The relying party and owner must fill and sign BudgetAdmission={host:Id, hardware:Digest, max_wall_seconds:U64, max_cost_minor:U64, currency:Id, max_memory_bytes:U64, max_disk_bytes:U64, threads:U64, jobs:U64, max_proof_seconds:U64, max_verify_seconds:U64, max_native_overhead_basis_points:U64, approved_by:Id, approval_ref:Digest} before any non-bootstrap campaign. All values are positive except a zero cost cap for explicitly free resources; currency uses a declared fixed unit.
INV-BUDGET: setup, compilation, witness generation, failed attempts, retries and teardown count toward the admitted cost/time cap. A lower-complexity graph needs a new candidate identity; weakening the verifier is never a performance optimization.
Measurement procedure¶
Freeze dataset/length roster, native baseline, exact model/graph/backend, setup reuse policy, host and repetitions before measurement. Use identical requests with confidential mode on/off where authorized, and state the engine, concurrency and exclusion rules. Record first-use setup separately; amortized setup cost states its reuse count. Preserve every failure, timeout and unsupported operator.
Report p50/p95 per stage, throughput with wall-clock denominator, peak resident/device memory, proof and bundle bytes, success fraction, and native release overhead relative to the matched baseline. A source/model hash or host inference timing is not a proving result. Raw StageSample records precede prose.
RISK: a useful encoder may exceed the approved proof budget. Owner @AbdelStark resolves by the v0.3 qualification gate using measured costs, not an inherited paper estimate. v1 stays blocked until a qualified encoder and budget exist.